Download Image Segmentation and Classification using Neural Network

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts
no text concepts found
Transcript
International Journal of Computer Science & Information Technology (IJCSIT) Vol 16, No 1, February 2024
IMAGE SEGMENTATION AND CLASSIFICATION
USING NEURAL NETWORK
Fatema Tuj Zohra, Rifa Tasfia Ratri, Shaheena Sultana, and Humayara Binte Rashid
Department of Computer Science and Engineering
Notre Dame University Bangladesh
Abstract. Image segmentation and classification tasks in computer vision have proven to be highly effective using neural networks, specifically Convolutional Neural Networks (CNNs). These tasks have numerous
practical applications, such as in medical imaging, autonomous driving, and surveillance. CNNs are capable
of learning complex features directly from images and achieving outstanding performance across several
datasets. In this work, we have utilized three different datasets to investigate the efficacy of various preprocessing and classification techniques in accurssedately segmenting and classifying different structures
within the MRI and natural images. We have utilized both sample gradient and Canny Edge Detection
methods for pre-processing, and K-means clustering have been applied to segment the images. Image
augmentation improves the size and diversity of datasets for training the models for image classification.
This work highlights transfer learning’s effectiveness in image classification using CNNs and VGG 16 that
provides insights into the selection of pre-trained models and hyper parameters for optimal performance.
We have proposed a comprehensive approach for image segmentation and classification, incorporating preprocessing techniques, the K-means algorithm for segmentation, and employing deep learning models such
as CNN and VGG 16 for classification.
Keywords: Convolutional Neural Network, VGG 16, Image Segmentation, K-means, Image Classification.
1
INTRODUCTION
In the world of artificial intelligence (AI) and computer science, computer vision is a
branch that focuses on giving computers the ability to interpret, process, and comprehend
visual data from the outside environment. It involves developing algorithms and techniques for processing and analyzing images and videos, and extracting meaningful insights
and information from them. A Convolutional Neural Network (CNN) is a type of deep
neural network made for processing and evaluating that has a data grid-like structure,
together with pictures or movies. Regarding computer vision, image identification is a
key task and CNNs have emerged as the most advanced technique for it. Convolutional
layers are used to extract information from images and fully connected layers are used to
produce predictions in CNNs [30]. CNN is a sort of neural network developed primarily
for image recognition tasks. Deep learning is another approach in machine learning that
focuses on teaching neural networks. This carries out tasks that require intelligence that is
comparable to that of a person, such as speech recognition, object recognition in pictures,
and language translation. Recent developments in deep learning have helped the medical
imaging industry detect many diseases [4]. To obtain high accuracy, the main purpose of
medical image classification is to identify the parts of the human body that are harmful
to health [5]. In the pre-processing stage, MRI images are preprocessed to remove noise
and enhance contrast using a combination of histogram equalization and median filtering
techniques.
In the segmentation stage K-means clustering algorithm divides the images into homogeneous segments based on color, texture, or intensity [29]. K-means is an unsupervised
learning algorithm that partitions data into K clusters, initially selecting random points
DOI: 10.5121/ijcsit.2024.16102
13
International Journal of Computer Science & Information Technology (IJCSIT) Vol 16, No 1, February 2024
as cluster centers and iteratively refining them. The algorithm assigns data points to the
nearest cluster center and updates centers based on the mean of points in each cluster.
The process continues until cluster centers stabilize or a specified iteration limit is reached.
Subtractive clustering is then employed to enhance segmentation by eliminating noise and
merging similar clusters, demonstrating the effectiveness through experimental results. In
the classification stage, pre-processed images are fed into a CNN model that is based
on the VGG 16 architecture [6]. Nath et.al. provided a comprehensive overview of various
image classification methods, including traditional techniques and deep learning-based approaches. They discussed the benefits and drawbacks of various approaches and how they
were applied in various fields [7].
The motivation behind employing neural networks, particularly CNNs, for image segmentation and classification lies in their demonstrated effectiveness in various computer vision
tasks. Neural networks offer improved accuracy, faster processing speeds, adaptability to
diverse data types, and the potential for novel applications. By leveraging the structure
and operation inspired by the human brain, these models can autonomously process and
analyze images, reducing the reliance on manual intervention.
However, the adoption of neural networks in image processing poses challenges. Training large-scale deep neural networks demand substantial computational resources, and the
scalability of such models remains a significant challenge. Additionally, ensuring generalizability across diverse datasets and real- world scenarios require addressing issues related to
overfitting and model robustness. Furthermore, interpretability and explainability of neural network decisions can be challenging, especially in critical applications where human
understanding is crucial. Balancing the trade-off between model complexity and computational efficiency is an ongoing challenge in the deployment of neural networks for image
segmentation and classification. Despite challenges related to interpretability and generalizability, the use of artificial intelligence-based visual systems, as demonstrated in fruit
classification through cameras and algorithms, showcase the potential for autonomous image analysis. Continued research and innovation in neural network techniques are essential
to overcome challenges and unlock the full potential of these models in revolutionizing image processing and analysis.
We have structured the rest of the paper as follows: Section 2 reviews the related works
in this field, highlighting the gaps in current knowledge and explaining how this work
addresses those gaps. Section 3 describes the models used for classification. Section 4 describes the methodology used to conduct the work. Section 5 presents the result of this
work, including performance analysis. Section 6 summarizes the findings of our work,
concludes, and suggests avenues for future research.
2
LITERATURE REVIEW
This section reviews the related works in this field, highlights the gaps in current knowledge, and explains how this work addresses those gaps. Methodological innovation is an
important contribution to method articles, and editors typically ask how the technique
in question differs from previously published methods. Image segmentation and classification is a supreme task in the execution of computer vision. Object recognition, medical
image analysis, autonomous driving, etc. are part of computer vision. Deep learning methods such as CNN, VGG 16, and k-means clustering have been admired in recent years
for these applications. We have organized our literature review into two distinct sections,
each addressing a specific aspect of our work: Convolutional Neural Networks (CNNs) and
Transfer Learning.
14
International Journal of Computer Science & Information Technology (IJCSIT) Vol 16, No 1, February 2024
In the section dedicated to Convolutional Neural Networks, we have extensively reviewed
prior research endeavors related to CNN models. Specifically, we have delved into the
historical body of work surrounding CNNs, analyzing their development and various applications.
In the context of Transfer Learning, our focus has been on VGG 16, a prominent
model in this domain. Within this section, we have incorporated relevant studies and
findings concerning VGG 16’s usage and its adaptations in the realm of transfer learning.
This approach enables us to comprehensively explore the landscape of previous research,
encompassing both the broader CNN field and the specific contributions of VGG 16 in
transfer learning applications.
2.1
Convolutional Neural Network (CNN)
CNNs have an architecture that helps them understand images step by step. They do this
by using three layers(convolutional layers, pooling layers, and fully connected layers) that
scan the image, and group information, and then make sense of it. AlexNet, VGGNet,
GoogLeNet, ResNet, and DenseNet are just a few of the CNN designs that Sultana et al.
offered an overview of, along with information on how well they performed on well-known
image classification benchmarks like ImageNet [8]. A CNN model with two completely
connected layers and four convolutional layers was produced by Khan et al. To avoid overfitting, they employed a dropout and Rectified Linear Unit (ReLU) activation function.
The rate of dropout is 0.5 [9]. A well-known benchmark dataset for computer vision is the
MNIST dataset. To attain high accuracy, Chattopadhyay et al. suggested a CNN architecture that combines convolutional, max-pooling, dropout, and fully connected layers with
optimal hyperparameters [10]. Kumar et al. proposed convolutional layers, which make up
the CNN’s architecture have various filter sizes and pooling layers for down-sampling the
feature maps. In addition, the authors employ batch normalization and dropout methods
to reduce overfitting and increase the network’s generalization capabilities [11]. As proposed by Kaushik et al. an approach that uses a CNN to learn the features of the image
and predict the segmentation map. Gomez et al. presented an approach that is evaluated
using a dataset of thermal images obtained from breast cancer patients and healthy subjects. According to the results, the suggested method works well in categorizing thermal
pictures into normal and malignant breast tissues, and it has the potential to be employed
as a non-invasive tool for early breast cancer testing [12]. Tripathi analyzed the effect of
different factors, such as network architecture, data augmentation, and hyperparameters,
on classification accuracy. They concluded that deeper networks with appropriate regularization techniques and data augmentation can significantly improve classification accuracy
[13].
2.2
Transfer Learning
Transfer learning is a strong approach that allows pre-trained models to be utilized for new
image categorization problems. It has proven cutting-edge performance on many benchmarks. Transfer learning is applied to a range of applications that also involve medical
picture classification. To identify brain tumors in MRI images, Siddique et al. proposed a
CNN model that has a high level of accuracy. The proposed model is used as a diagnostic
tool in clinical settings to aid radiologists in the detection of brain tumors [14]. Agarwal
proposed a deep-learning approach for classifying cooking images into different states using
the VGG 19 network [15]. D.C. Febrianto et al. suggested a method for training a CNN
15
International Journal of Computer Science & Information Technology (IJCSIT) Vol 16, No 1, February 2024
using a dataset of brain magnetic resonance imaging (MRI) images that contain both normal and tumor scans [16]. Hoque provided valuable insights into the application of deep
learning models for medical image analysis, specifically for brain tumor detection, and
highlighted the potential benefits of using CNNs for this task. The comparative analysis
of VGG 16 and VGG 19 models provides a useful benchmark for future studies in this area
[17]. Abd-Ellahet al. compared the performance of the VGG 16 and VGG 19 networks on a
dataset of 500 MRI scans, consisting of 250 normal scans and 250 scans with tumors. They
also compared the performance of the VGG networks with a conventional CNN approach
[18]. Pravallika and Baskar suggested a brain tumor classification method based on image
processing that employs the VGG 16 CNN and classifiers using support vector machines
(SVM). Analysis of the proposed system on a dataset of 210 brain MRI images, and the
results show that the VGG 16 network outperforms the SVM classifier with an accuracy
of 95.2% compared to 89.5% [19]. Agus et al. proposed a system that involves training the
VGG 16 model on the MRI images to extract features followed by a classification layer
using softmax regression for classifying the image into one of the two categories, glioma
or non glioma [20]. Simonyan et al. introduced the VGGNet architecture, which achieved
outstanding performance on the ImageNet dataset. They also demonstrated the efficiency
of transfer learning for image classification by adjusting the previously trained VGGNet
on a smaller dataset [34]. Long et al. presented a deep adaptation network (DAN) that can
learn domainagnostic characteristics. They used various photos to highlight the usefulness
of DAN.
3
MODELS USED IN IMAGE CLASSIFICATION
In this section, we have discussed the models that have been used for classification. We
have used CNN and VGG 16 to classify our images.
3.1
Convolutional Neural Network
A convolutional neural network has three layers. Input layer, hidden layer, and output
layer. Data is received by the input layer, then it is sent to hidden layers. The process of
removing features from the data is carried out by the hidden layers, and each layer has
multiple nodes or neurons that perform the calculations. Based on the features that The
output layer has been gathered by the hidden layers generates the outcome or prediction.
At the time of training, the weights of the nodes are adjusted. This has been done to
minimize the error between the predicted output and the actual output. This process has
been repeated iteratively until the network achieves the desired accuracy. CNNs, in particular, are specialized neural networks that have been designed to process images and other
types of multidimensional data [30]. Convolutional Neural Networks are often employed in
image categorization, object identification, and other computer vision applications. CNNs
are built to automatically learn and extract features from pictures using convolution and
pooling. Convolution is performed by applying a tiny filter or kernel over an input picture
and computing the dot product of the filter with each patch of pixel then the output is
routed via an activation function like ReLU to induce nonlinearity. The feature maps are
then down-sampled and the output dimensionality is reduced via pooling. Figure 1 shows
us the architecture of convolutional neural network [32]. Here are the key details of CNN:
1. Convolutional Layer: To extract characteristics from the input image, this layer has
used several filters (kernels). The filters convolve over the image and produce a feature
map. Convolutional layers can learn low-level features like edges, lines, and curves. The
16
International Journal of Computer Science & Information Technology (IJCSIT) Vol 16, No 1, February 2024
formula for the 2D convolution operation in a convolutional layer can be expressed as
follows:
2. ReLU Activation Layer: ReLU is an activation function for Rectified Linear Units.
This activation function has been applied to the convolutional layer’s output. This aids
in introducing non-linearity into the model and permits it to pick up on more intricate
aspects. The ReLU function could be defined as:
f (x) = max(0, x)
(1)
3. Pooling Layer: This layer has been used to reduce the size of the feature maps. It
has taken the maximum or average value of each patch of pixels. This has decreased
the size of feature maps which helped to reduce overfitting.
4. Fully Connected Layer: It has taken the output of the layer and runs it through
a group of neurons that are completely linked to every neuron in the layer before. To
categorize the supplied image, utilize this layer.
5. Softmax Layer: For generating a probability distribution over the classes, it has
applied a softmax function to the output of the fully connected layer. The softmax
function follow as:
e xi
softmax(xi ) = P xj
(2)
je
6. Loss Function: A loss function computes the differentiation between two values. a
model’s anticipated output and the actual result (i.e., the ground truth) for a certain input. It is a crucial component of building a machine-learning model because it
has directed the optimization process to the best possible collection of model parameters. CNNs can be trained using a variety of optimization algorithms, such as SDG
(Stochastic gradient descent) or Adam. They can also be trained on large datasets,
such as ImageNet, and fine-tuned on smaller datasets for specific tasks. The most advanced performance on a wide range of computer vision tasks is possible with CNNs
when trained via backpropagation and gradient descent. Numerous more uses include
style transfer, segmentation, object identification, and image classification.
Fig. 1: CNN Architecture.
3.2
Visual Geometry Group
VGG (Visual Geometry Group) is a group of researchers from Oxford University that
specializes in computer vision and deep learning. They are well-known for their contributions to the area of convolutional neural networks that excel in image recognition tasks.
17
International Journal of Computer Science & Information Technology (IJCSIT) Vol 16, No 1, February 2024
In 2014, Simonyan et al. suggested a new architecture for CNNs that delivered outstanding outcomes on several image recognition benchmarks [34]. This architecture is known
as VGGNet. This consists of a series of convolutional layers with small 3x3 filters. These
layers are followed by max-pooling layers and end with several fully connected layers. The
VGGNet is known for its simplicity and is still widely used today as a baseline for image
recognition tasks. VGG 16 deep neural network with 16 layers. Figure 2 shows us how the
VGG 16 model works layer by layer [33]. Here are the details of the VGG 16 architecture:
1. Input Layer: A RGB image with the dimensions 224 x 224 x 3 have been used as the
input layer.
2. Convolutional Layers: There are 13 convolutional layers in VGG 16. Each convolutional layer has a 3 x 3 kernel and used a stride of 1 pixel. Each layer’s number of
filters increased as it gone deeper into the network, starting with 64 filters in the first
layer and doubling after each max pooling layer.
3. Max Pooling Layers: There are 5 max-pooling layers in VGG 16. Each max-pooling
layer has a 2 x 2 kernel and a stride of 2 pixels.
4. Fully Connected Layers: VGG 16 has 3 layers that have been completely linked.
The amount of neurons in the first two totally connected layers (4,096 neurons each)
and the third fully connected layer (1,000 neurons) correspond to the number of classes
in the ImageNet dataset.
5. Softmax Layer: For obtaining the final probability distribution across the 1,000
classes, a softmax function have applied to the output of the last fully connected
layer.
Fig. 2: VGG 16 Architecture.
4
METHODOLOGY
This section describes the methodology used to conduct the work. In our work, we have
followed a procedure that contains all of the methods or tools that we have used to analyze
image segmentation and classification in our work. Figure 3 shows the procedure of our
work. First of all, we have collected a dataset. We have worked on three datasets. All of
them have been collected from Kaggle [1].
18
International Journal of Computer Science & Information Technology (IJCSIT) Vol 16, No 1, February 2024
Fig. 3: Workflow of the work
4.1
Data Collection
In our work, we have chosen three datasets shown in Table 1 from Kaggle based on two
classes of training and testing [1]. The first and second datasets are about brain tumor
images [21] and [22]. Dataset 1 and Dataset 2 contain around 3274 images and 7023 images
in total. They have 4 classes including glioma tumors, no tumors, meningioma tumors,
and pituitary tumors. The third dataset is about natural images. This dataset contains a
collection of 7 categories of natural images with a total of 8,789 images [1]. In our work,
we select four categories such as dog, cat, fruit and flower.
Splitting Category
Dataset 1 Dataset 2 Dataset 3
Train Test Train Test Train Test
2880 394 5712 1311 2642 788
Table 1: Splitting Category of all Datasets
4.2
Dataset Preprocessing
Before analyzing the dataset, it is necessary to preprocess the dataset for better results.
The goal of preprocessing is to prepare the data for analysis and to increase the model’s
accuracy by reducing noise and removing inconsistencies. We have used Canny Edge Detection and Sample Gradient as preprocessing steps.
Canny Edge Detection: This is a widely used algorithm for detecting edges in images. It worked by identifying areas in the image with a significant change in intensity or
color and marking them as edges. The algorithm involves several steps, including smoothing the image to remove noise and calculating the gradient of the image to determine the
edges, and non-maximum suppression is used to soften the edges [31]. The edges serve as
the final threshold to create a binary picture, with the background represented by black
19
International Journal of Computer Science & Information Technology (IJCSIT) Vol 16, No 1, February 2024
pixels and the edges by white pixels. Figure 4 shows us how canny edge detection works
on four types of MRI images and natural images.
(a) MRI Images
(b) Natural Images
Fig. 4: Canny Edge Detection
Sample Gradient: In image processing, the gradient is a measure of the rate at which
the pixel intensity changes in an image. In our work, we have used the Sobel gradient
for image preprocessing. The Sobel gradient is a simple and widely used method for edge
detection in image processing. S. B. Kulkarni and S. G. Bhirud described the Sobel gradient
as a simple and effective method for edge detection. They explain that the Sobel gradient
works by calculating the gradient magnitude of an image by convolving it with two filters,
one for the horizontal edges and another for the vertical edges [23]. Mustafa et al. explained
that the Sobel gradient is a popular edge detection method due to its simplicity and
effectiveness [24]. Wang proposed work to develop Laplacian operator-based edge detectors.
The detectors utilized the second-order derivative of Gaussian filters to extract the edges
from an image [30].
4.3
Image Segmentation
The technique of segmenting an image involves dividing it up into different sections or areas
that are uniform in terms of things like color, texture, or intensity. Minaee et al. explain
the importance of image segmentation in different uses of computer vision including object
identification, recognition, and classification [25]. FCNs are a kind of deep neural network
that can segment images pixel by pixel. To learn how to divide an image into several areas
according to their semantic significance, Long et al. suggested an approach that involves
training an FCN on a sizable dataset of annotated images [26]. One way to perform image
segmentation is using the K-means. Data are divided into K clusters based on similarity
using the unsupervised learning method K-means. Burney et al. presented a clear and
short overview of the application of K-means clustering for segmentation. It highlights
the advantages of this approach and demonstrates its effectiveness through experimental
results [27].
4.4
Data Augmentation
Augmentation is used to boost the quantity and variety of training data without actually
gathering any new data. By applying various transformations to existing data, such as
rotating, cropping, or adding noise, it has created a new sample that is still representative
of the original data for MRI images and natural images as shown in figure 5. The goal
of data augmentation is to increase the diversity and volume of training data, which can
improve deep learning models’ generalization performance [28].
20
International Journal of Computer Science & Information Technology (IJCSIT) Vol 16, No 1, February 2024
(a) MRI Images
(b) Natural Images
Fig. 5: Data Augmentation
4.5
Our CNN Model Structure
In CNN models, various layers work together in a structured manner. Convolutional layers
extract image features, pooling layers decrease spatial resolution for efficiency, and fully
connected layers handle the final classification or regression task, depending on the problem
at hand. This is a sequential neural network model with the following layers:
1. Conv2D layer with 32 filters, kernel size of (3,3), ReLU activation function, and input
shape of (IMAGE SIZE, IMAGE SIZE, 3). This layer applies 32 different filters to the
input image, each of size 3x3, and applies the ReLU activation function to the output.
2. MaxPooling2D layer with pool size of (2,2). This layer reduces the size of the input
image by taking the maximum value within each 2x2 window.
3. Conv2D layer with 64 filters, kernel size of (3,3), and ReLU activation function. This
layer applies 64 different filters to the output of the previous layer, each of size 3x3,
and applies the ReLU activation function to the output.
4. MaxPooling2D layer with pool size of (2,2). This layer reduces the size of the input
image by taking the maximum value within each 2x2 window.
5. Conv2D layer with 32 filters, kernel size of (3,3), and ReLU activation function. This
layer applies 32 different filters to the output of the previous layer, each of size 3x3,
and applies the ReLU activation function to the output.
6. Flatten layer. This layer flattens the output of the previous layer into a 1D array.
7. Dense layer with 16 neurons and ReLU activation function. This layer applies a fully
connected layer to the output of the previous layer, with 16 neurons and ReLU activation function.
8. Dense layer with 4 neurons and softmax activation function. This layer applies a fully
connected layer to the output of the previous layer, with 4 neurons and softmax activation function.
The model have compiled using the ’adam’ optimizer, ’sparse categorical crossentropy’
as the loss function, and ’accuracy’ as the evaluation metric.
4.6
Our Transfer Learning Model
Transfer learning approach entails leveraging previously learned models as a foundation for
new models. The VGG 16 architecture is a popular choice for transfer learning due to its
excellent performance in image recognition tasks. We have built a transfer learning model
based on the VGG 16 architecture using the Keras API. The pre-trained VGG 16 model
is loaded with the imagenet weights, and all the layers are set to non-trainable, except for
the last three layers in the VGG block, which are set to trainable. The pre-trained VGG
21
International Journal of Computer Science & Information Technology (IJCSIT) Vol 16, No 1, February 2024
16 model comes first followed by a flattened layer, two dropout layers, a dense layer with
128 neurons with ReLU activation, and a final dense layer with neurons. The neurons are
equal to the number of unique labels in the dataset and a softmax activation function.
This new sequential model is then defined. The model has been made for tasks requiring
several classes of categorization.
5
RESULT ANALYSIS
This section presents the result of this work, including performance analysis and shows
how Convolutional Neural Networks (CNN) and VGG 16 architecture classify our different
datasets into different categories, focusing on evaluating their performance.
This analysis used CNN and VGG 16 for classifying our datasets into different categories. The Adam optimizer have been used to train the neural network. We have used
10 epochs, a batch volume of 20, and sparse categorical cross-entropy as the loss function.
The objective was to evaluate the performance of these models in accurately classifying
MRI images and comparing their results. We have reported the results for 3 datasets which
shows that the accuracy of VGG 16 is better than CNN for all datasets. For dataset 1,
the CNN model training accuracy is 99% and test accuracy is 72%, while the VGG 16
achieved 98% in the training and 75% in the testing phase. In dataset 2, the CNN model
achieved and VGG 16 both models achieved 99% in the training phase, but in the testing
phase VGG 16 model result is better than the CNN model. In dataset 3, the CNN model
achieved a training accuracy of 99% and a test accuracy of 82%. But VGG 16 model
achieved 99% in the training phase and 95% in the testing. Table 2 shows the accuracy
performance of the training and test sample for all models.
Model
Dataset 1
Train Test
CNN 99
72
VGG16 98
75
Dataset 2
Train Test
99
95
99
97
Dataset 3
Train Test
99
82
99
95
Table 2: Accuracy performance between CNN and VGG 16
5.1
Evaluation Matrix
In machine learning, performance metrics are used to evaluate the effectiveness of a model.
One such metric is the classification report, which evaluates the quality of a classification
model. The report includes various performance metrics such as accuracy, precision, recall,
F1-score, and support [2].
– True Positive(TP): The model successfully predicts the positive class and the actual
outcome is positive.
– True Negative(TN): The model successfully predicts the negative class when the
actual outcome is truly negative.
– False Positive(FP): The model predicts the positive class inaccurately when the real
class is negative.
– False Negative(FN): The model predicts the negative class inaccurately when the
real class is positive.
22
International Journal of Computer Science & Information Technology (IJCSIT) Vol 16, No 1, February 2024
– Accuracy:Accuracy measures the percentage of correct predictions made by the model
Accuracy = T P + T N/P + N
(3)
– Precision: Precision measures the per centage of positive images that were correctly
classified out of all the images that the model predicted as positive.
P recision = T P/T P + F P
(4)
– Recall:Recall measures the percentage of positive images that were correctly classified
out of all the actual positive images.
Recall = T P/T P + F N
(5)
– F1-Score:The F1-score is the harmonic mean of precision and recall and is used to
evaluate the overall performance of the model.
F 1score = 2 ∗ (P recision ∗ Recall)/(P recision + Recall)
(6)
– Support: Support is the total number of images in a certain class. It represents the
number of instances in the dataset that belong to a particular class.
By evaluating these metrics, researchers can determine the quality of their model and
how well it has being performing in accurately identifying the correct labels for a given set
of images. Finally, we have calculated the Precision, Recall, and F1-Score of the proposed
CNN and VGG 16 model, which are commonly used performance evaluation metrics to
measure the effectiveness of an image classification model in accurately identifying the
correct labels for a given set of images.
In this work, experiments are conducted using three distinct datasets to train Convolutional Neural Network (CNN) and VGG16 models, varying in the sizes of training and test
samples. Notably, Dataset 1, with 3000 data points, exhibited lower F1 scores compared
to the other two datasets. The decision to exclude results from Dataset 1 is based on multiple considerations. Firstly, the relatively small size of the dataset may limit the models’
ability to generalize effectively, potentially leading to decrease performance. Additionally,
concerns about an imbalanced distribution of samples across classes in Dataset 1 could
bias the models, particularly affecting performance metrics like the F1 score, especially
for minority classes. In conclusion, the exclusion of results from Dataset 1 is a strategic
choice aimed at ensuring the robustness and reliability of the conclusions derived from the
experiments.
Convolutional Neural Network
Precision Recall F1 score Support
no tumor
1.00
0.99
0.99
405
meningioma tumor 0.94
0.87
0.90
306
glioma tumor
0.90
0.96
0.93
300
pituitary tumor
0.97
0.98
0.98
300
Convolutional Neural Network
Precision Recall F1 score Support
cat
0.81
0.67
0.73
259
dog
0.58
0.67
0.62
159
flower 0.86
0.93
0.90
183
fruit
0.97
0.99
0.98
187
(a) Dataset 2
(b) Dataset 3
Table 3: Evaluation Matrix using CNN
23
International Journal of Computer Science & Information Technology (IJCSIT) Vol 16, No 1, February 2024
The accuracy, recall, F1 score, and support for the CNN using Dataset 2 and Dataset
3 are shown in Table 3. The table demonstrates that Dataset 3 have a lower precision,
recall, and F1 scores than Dataset 2. Dataset 2 shown in Table 3a has the best precision,
recall, and F1 score.
VGG 16
Precision Recall F1 score Support
glioma tumor
0.94
0.98
0.96
300
meningioma tumor 0.97
0.94
0.96
306
no tumor
1.00
1.00
1.00
405
pituitary tumor
0.99
0.98
0.99
300
(a) Dataset 2
cat
dog
flower
fruit
Precision
0.92
0.96
0.98
1.00
VGG 16
Recall F1 score Support
0.98
0.95
259
0.84
0.90
159
0.99
0.99
183
1.00
1.00
187
(b) Dataset 3
Table 4: Evaluation Matrix using VGG 16
Table 4 displays the Precision, Recall, F1 score, and Support results for the VGG 16
using Dataset 2 and Dataset 3. The table demonstrates that Dtaset 2 has the highest
F1 score, recall, or accuracy than Dataset 3. For Dataset 2 shown in Table 4a the model
performs flawlessly for the assigned task, as seen by its 100% accuracy, 100% recall, and
100% F1 score for no tumor. This would suggest that there were no false positives or
false negative predictions made by the model, meaning that it properly categorised every
incident in the dataset.
(a) CNN
(b) VGG 16
Fig. 6: Model Training History of Dataset 2
Figure 6 is a training graph of accuracy and loss for CNN and VGG 16 that shows
the performance of the model throughout training for Dataset 2. The accuracy graph of
CNN in figure 6a shows how the accuracy of the model changes over time during training.
Initially, the accuracy may be low, but as the model is trained on more data, it gradually
improves to 99.99%. The loss graph shows how the loss of the model changes over time
during training. The loss should generally decrease over time, as the model learns to make
better predictions.
The accuracy graph for VGG 16 in figure 6b would show the percentage of images
that were correctly classified by the model. Initially, the accuracy would be low, but as
the model is trained on more data, the accuracy would increase by 99.99%. The loss graph
24
International Journal of Computer Science & Information Technology (IJCSIT) Vol 16, No 1, February 2024
for VGG 16 shows the amount of error between the predicted and actual labels for the
training data. The loss should generally decrease over time as the model learns to make
better predictions.
(a) CNN
(b) VGG 16
Fig. 7: Model Training History of Dataset 3
Figure 7 is a training graph of accuracy and loss for CNN and VGG 16 that shows
the performance of the model throughout training for Dataset 3. The accuracy graph of
CNN in figure 7a shows how the accuracy of the model changes over time during training.
Initially, the accuracy may be low, but as the model is trained on more data, it gradually
improves to 99%. The loss graph shows how the loss of the model changes over time during
training. The loss should generally decrease over time, as the model learns to make better
predictions.
The accuracy graph for VGG 16 in figure 7b would show the percentage of images that
were correctly classified by the model. Initially, the accuracy would be low, but as the
model is trained on more data, the accuracy would increase by 99.99%. The loss graph
for VGG 16 shows the amount of error between the predicted and actual labels for the
training data. The loss should generally decrease over time as the model learns to make
better predictions.
6
CONCLUSION
This study presents a comprehensive approach for image segmentation and classification
using pre-processing techniques, K-means algorithm for segmentation, and deep learning
models (CNN and VGG16) for classification. Higher accuracy both for MRI and natural
images is achieved by proposed approach. The VGG 16 model outperforms than the other
models in terms of accuracy. It is shown that pre-processing techniques, segmentation, and
deep learning models can be combined to achieve high accuracy in image classification
tasks. The proposed method has the potential to be used in a variety of medical conditions, as MRI imaging is commonly used in medical diagnosis and treatment planning. In
future research we can explore larger datasets, alternative segmentation techniques, transfer learning, ensemble methods, and alternative metrics. These techniques can be used to
improve the robustness, accuracy, and applicability of the system. Further research can
explore the deployment of these systems in clinical settings and real-time segmentation
on low-power devices. As the field continues to evolve, there will be exciting opportunities
for more advanced techniques for analyzing and understanding visual data.
References
1. Kaggle, https://www.kaggle.com/, [Online; accessed 22-March-2023]
25
International Journal of Computer Science & Information Technology (IJCSIT) Vol 16, No 1, February 2024
2. “Javatpoint.” https://www.javatpoint.com/performance-metrics-in-machine-learning [Online; accessed
25 December-2023].
3. Roy, Prasun,Natural Images, year = 2021, https://www.kaggle.com/prasunroy/natural-images, [Accessed: March 23, 2023]
4. N. Abiwinanda, M. Hanif, S. T. Hesaputra, A. Handayani, and T. R. Mengko, “Brain tumor classification
using convolutional neural network,” in World Congress on Medical Physics and Biomedical Engineering
2018: June 3-8, 2018, Prague, Czech Republic (Vol. 1), pp. 183–189, Springer, 2019
5. E. Miranda, M. Aryuni, and E. Irwansyah, “A survey of medical image classification tech- niques,”
in 2016 International Conference on Information Management and Technology (ICIMTech), pp. 56-61,
2016
6. B. Padmini, C. Johnson, B. A. Kumar, and G. R. Yadav, “Brain tumor detection by using cnn and
vgg-16,”.
7. S. S. Nath, G. Mishra, J. Kar, S. Chakraborty, and N. Dey, “A survey of image classification methods
and techniques,” in 2014 International Conference on Control, Instrumentation, Communication and
Computational Technologies (ICCICCT), pp. 554-557, 2014.
8. F. Sultana, A. Sufian, and P. Dutta, “Advancements in image classification using convolu- tional neural
network,” CoRR, vol. abs/1905.03288, 2019.
9. H. A. Khan, W. Jue, M. Mushtaq, and M. U. Mushtaq, “Brain tumor classification in mri image using
convolutional neural network,” Mathematical Biosciences and Engineering, vol. 17, no. 5, pp. 6203-6216,
2020.
10. N. Chattyopadhyay, H. Maity, B. Debnath, D. Ghosh, R. Chakraborty, S. Mitra, R. Islam, and N.
Saha, “Classification of mnist image dataset using improved convolutional neural network”.
11. P.Kumar and U.Dugal, “Tensorflow based image classification using advanced convo- lutional neural
network,” International Journal of Recent Technology and Engineering (IJRTE), vol. 8, pp. 994-998,
2020.
12. J. Zuluaga-Gomez, Z. Al Masry, K. Benaggoune, S. Meraghni, and N. Zerhouni, “A cnn-based methodology for breast cancer diagnosis using thermal images,” Computer Methods in Biomechanics and
Biomedical Engineering: Imaging & Visualization, vol. 9, no. 2, pp. 131-145, 2021.
13. M. Tripathi, “Analysis of convolutional neural network based image classification tech- niques,” Journal
of Innovative Image Processing (JIIP), vol. 3, no. 02, pp. 100-117, 2021.
14. M. A. B. Siddique, S. Sakib, M. M. R. Khan, A. K. Tanzeem, M. Chowdhury, and N. Yas- min,
“Deep convolutional neural networks model-based brain tumor detection in brain mri images,” in 2020
Fourth International Conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud)(I-SMAC),
pp. 909-914, IEEE, 2020.
15. R. Agarwal, “State classification of cooking images using vgg19 network”.
16. D. Febrianto, I. Soesanti, and H. Nugroho, “Convolutional neural network for brain tu- mor detection,”
in IOP Conference Series: Materials Science and Engineering, vol. 771, p. 012-031, IOP Publishing, 2020.
17. S. U. Hoque, “Performance comparison between vgg16 & vgg19 deep learning method with cnn for
brain tumor detection”.
18. M. K. Abd-Ellah, A. I. Awad, A. A. Khalaf, and H. F. Hamed, “Two-phase multi-model automatic
brain tumour diagnosis system from magnetic resonance images using convo- lutional neural networks,”
EURASIP Journal on Image and Video Processing, vol. 2018, no. 1, pp. 1-10, 2018.
19. C. R. Pravallika and R. Baskar, “Image processing based brain tumor classification using vgg16 compared with svm to improve accuracy,” pp. 1398-1401, 2022.
20. M. E. Agus, S. Y. Bagas, M. Yuda, N. A. Hanung, and Z. Ibrahim, “Convolutional neural network
featuring vgg-16 model for glioma classification,” JOIV: International Journal on Informatics Visualization, vol. 6, no. 3, pp. 660-666, 2022.
21. S. Bhuvaji, A. Kadam, P. Bhumkar, S. Dedge, and S. Kanchan, “Brain tumor classification (mri),”
2020.
22. M. Nickparvar, “Brain tumor mri dataset,” 2021.
23. S. B. Kulkarni and S. G. Bhirud, “Edge detection techniques for image segmentation: a survey of soft
computing approaches,” International Journal of Emerging Technology and Advanced Engineering, vol.
4, no. 8, pp. 610-614, 2014.
24. Z. Mustafa, M. A. Khan, and M. Riaz, “An overview of image edge detection techniques,” Journal of
Information Processing Systems, vol. 13, no. 1, pp. 1-20, 2017.
25. S. Minaee, Y. Boykov, F. Porikli, A. Plaza, N. Kehtarnavaz, and D. Terzopoulos, “Image segmentation
using deep learning: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.
44, no. 7, pp. 3523-3542, 2022.
26. . Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic seg- mentation,” in
Proceedings of the IEEE conference on computer vision and pattern recog- nition, pp. 3431-3440, 2015.
27. S. A. Burney and H. Tariq, “K-means cluster analysis for image segmentation,” Interna- tional Journal
of Computer Applications, vol. 96, no. 4, 2014.
26
International Journal of Computer Science & Information Technology (IJCSIT) Vol 16, No 1, February 2024
28. L . Perez and J. Wang, “The effectiveness of data augmentation in image classification using deep
learning,” CoRR, vol. abs/1712.04621, 2017.
29. N. Dhanachandra, K. Manglem, and Y. J. Chanu, “Image segmentation using k-means clustering
algorithm and subtractive clustering algorithm,” Procedia Computer Science, vol. 54, pp. 764–771,
2015.
30. X. Wang, “Laplacian operator-based edge detectors,” IEEE transactions on pattern analysis and machine intelligence, vol. 29, pp. 886–90, 06 2007.
31. Sekehravani, E. A., Babulak, E., and Masoodi, M. “Implementing canny edge detection algorithm for
noisy image. Bulletin of Electrical Engineering and Informatics”, 9(4), 1404-1410,2020.
32. R. Kaushik and S. Kumar, “Image segmentation using convolutional neural network,” Int. J. Sci.
Technol. Res, vol. 8, no. 11, pp. 667–675, 2019.
33. M. E. Agus, S. Y. Bagas, M. Yuda, N. A. Hanung, and Z. Ibrahim, “Convolutional neural network
featuring vgg-16 model for glioma classification,” JOIV: International Journal on Informatics Visualization, vol. 6, no. 3, pp. 660–666, 2022.
34. K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,”
arXiv preprint arXiv:1409.1556, 2014.
Authors
Fatema Tuj Zohra is a recent B.Sc. graduate in Computer Science and Engineering. With
a passion for advancing the field of machine learning, Fatema combines a strong foundation in computer science with a profound curiosity for exploring innovative solutions. She
seeks to revolutionize real-world applications through their insightful contributions and
unwavering dedication to technological progress. .
Rifa Tasfia Ratri received B.Sc. Engineering degree in Computer Science and Engineering (CSE) from Notre Dame University Bangladesh. Her current research focuses on
machine learning and image processing.
Dr. Shaheena Sultana is a Professor and Chairman in the department of Computer
Science and Engineering at Notre Dame University Bangladesh. She received M.Sc. Engineering degree in Computer Science and Engineering (CSE) from Bangladesh University
of Engineering and Technology (BUET). She did B.Sc. Engineering degree in Electrical
and Electronic Engineering (EEE) from Khulna University of Engineering and Technology
(KUET) She obtained Ph.D. in CSE from BUET. Her research interests include Graph
Drawing, Graph Theory, VLSI Design, Embedded System, Data Mining.
Humayara Binte Rashid is currently teaching as a Lecturer in the Department of
Computer Science and Engineering of Notre Dame University Bangladesh (NDUB). She
passed B.Sc. in Computer Science and Engineering (CSE) from Military Institute of Science and Technology (MIST). Her current research focuses on machine learning and Data
Mining.
27